The debate about artificial intelligence still revolves obsessively around models, parameters and chips. But the real bottleneck that will determine who leads this race is not computational — it is physical. Training a model on the scale of GPT-4 consumes somewhere between 50 and 100 gigawatt-hours of electricity, equivalent to the annual consumption of tens of thousands of homes. Doing inference at scale, serving billions of daily queries, multiplies this number by orders of magnitude. The strategic question is not who has the best algorithms. It's who has access to cheap, abundant and reliable energy — and for how long.
The electrical cost that roadmaps ignore
Training frontier models is a discrete operation: it happens once, lasts weeks or months, and the energy cost, although high, is predictable. The real problem is continuous inference. Every time a user sends a message to an AI assistant, processes a document, or generates an image, an energy-consuming cycle of computation takes place. Multiply that by hundreds of millions of simultaneous users and you have electrical demand that rivals entire industrial sectors.
Microsoft has projected that its AI commitments will triple the energy consumption of its data centers by the end of this decade. OpenAI, in partnership with SoftBank on the Stargate project, announced investments of 500 billion dollars in American infrastructure — and a large part of this amount is directly related to energy generation and transmission. Goldman Sachs has estimated that global data center demand is expected to grow 160% by 2030, driven primarily by AI payloads. These are not technology numbers. These are heavy infrastructure numbers.
The geopolitical map that is being redrawn
The distribution of cheap, clean energy across the planet is uneven, and this is creating structural advantages that few strategists are seriously mapping. The Nordic countries—Iceland, Norway, Sweden—combine abundant geothermal and hydropower with cool weather, which reduces data center cooling costs. France operates 70% of its matrix with nuclear energy, which guarantees stable prices and low emissions. Canada has huge hydroelectric plants concentrated in the provinces of Quebec and British Columbia. All of these countries today have an infrastructure advantage that cannot be quickly replicated by competitors.
On the opposite side, countries heavily dependent on imported fossil fuels face a double problem: high cost and supply instability. Building AI data centers in regions where electricity is expensive or prone to outages is not just inefficient — it's strategically risky. The geopolitics of energy has always shaped the geopolitics of power. What has changed is that it now also shapes the geopolitics of intelligence.
The nuclear renaissance and what it reveals
Microsoft's decision to reactivate the Three Mile Island plant in Pennsylvania, closed since 2019, was not nostalgia. It was engineering calculation. The plant will supply power exclusively to Microsoft data centers under a long-term contract. Amazon made a similar move by acquiring a data center campus directly connected to a nuclear plant in Pennsylvania. Google announced contracts with modular reactor company Kairos Power to supply 500 megawatts from 2030.
What these decisions reveal is simple: Big tech companies do not believe that solar and wind, in their current configuration, are capable of delivering the reliable baseload that AI demands. Solar and wind energy are intermittent — the sun doesn't shine at night and the wind doesn't always blow. AI data centers require 99.9999% uptime, without interruptions. Nuclear delivers exactly that: high generation density, predictability, and zero carbon emissions. What seemed like a technology in decline suddenly became a strategic asset of the highest order.
The Brazilian position: real advantage with problematic concentration
Brazil operates with one of the cleanest electrical matrices in the world. Approximately 85% of national electrical generation comes from renewable sources — hydroelectric, wind and solar. This puts the country in a privileged position in any scenario where the environmental cost of computing becomes a competitive factor, whether due to European regulation or pressure from corporate clients with ESG goals. A data center operating with 100% renewable energy in Brazil is not green marketing — it is a differentiated commercial proposition.
The expansion of wind in the Northeast and solar in the Cerrado has added significant capacity in recent years at prices that are among the most competitive in the world in energy auctions. The average cost of contracted wind energy in Brazil is consistently below R$150 per megawatt-hour, a number that most European markets cannot approach. The advantage exists. The problem is concentration: the country's data center corridor is disproportionately concentrated on the São Paulo-Rio axis, a region that depends on an old transmission network and reservoirs that suffer from climate variations. The structural advantage of the clean matrix is partially neutralized by the geography of concentrated consumption.
What does this change for those who decide on AI infrastructure
Organizations that still treat cloud region selection as a secondary technical decision — latency, regulatory compliance, service availability — need to broaden the analytical framework. Energy is entering the total cost of operation equation in a way that has no historical precedent in cloud computing. AWS, Azure and Google Cloud already differentiate prices between regions in part because of local energy costs. This difference will increase, not decrease.
For companies with AI operations at scale — whether in-house inference or intensive use of APIs — the localization of workloads has become a relevant cost lever. Running inference in regions with cheaper and cleaner energy reduces operational costs and regulatory exposure. Companies that are building their own infrastructure need to include the following dimensions in the location selection process: the region's electrical generation mix, grid stability, projected energy costs for five and ten years, and availability of long-term renewable energy purchase agreements. These are not secondary variables. For those who will operate data centers for decades, these are primary variables.
Brazil has a concrete opportunity to position itself as an AI data center destination for Latin American and global companies in need of cheap, renewable energy. But this requires investment in transmission to decentralize consumption, and industrial policy that encourages data centers outside the South-Southeast axis. The advantage exists on paper. Turning it into a real advantage requires political decision and infrastructure — two resources that, historically, the country takes longer to mobilize than it should.
Also read
- The AI energy crisis: what data center consumption means for infrastructure decision makers
- AI Energy and Sovereignty: What Governments Need to Plan Now
- Energy: The Bottleneck That Nobody Put on the Computing Roadmap
- Sustainable data centers: the environmental cost that became a public issue
- Electrification: what changes when everything starts to turn on socket
- Silicon geopolitics: why the chip war matters for those who decide about technology